As an initial proposal, the planning and management levels are intended to be performed by AI-based learning systems as well. Due to the large number of functions that are executed at these levels, for now, it is only considered to work with the functions deployed by the ERP: Order Processing (OPR) and Production Scheduling (PS), and MES: functions related to Control of Production (PC), as indicated in Fig. 2.3. Decision-making systems, such as MES and ERP, can be considered non-linear systems since they involve a considerable variety of scenarios and combinations due to the number of inputs, outputs, relationships and conditions that are carried out in the functions that are performed at these levels.
Usually, AI-based learning algorithms such as Multilayer Neural Networks (MNNs) are used as alternative to learn linear and non-linear relationships between input and output vec- tors. This is one of the main reasons for the use of MNNs. As is known, these algorithms can carry out operations with complex non-linear relationships between dependent and inde- pendent variables, such as the functions that are executed in the ERP and MES. In addition, MNNs offer advantages such as the ability to detect all possible interactions between predictor
variables, the availability of multiple training algorithms, the ability to work with incomplete knowledge, fault tolerance (in one or more cells) , parallel processing capacity, etc [59], [60].
MNNs learn events and make decisions from the examples (database) with which they are trained. All the characteristics mentioned are adjusted to the capabilities that this proposal seeks to provide to virtual decision-making systems based on the DT. Further information about Artificial Neural Networks and particularly Back-Propagation Networks used in this work, will be discussed in Chapter 5.
MNN algorithm consists of a system of interconnected neurons, or nodes. The neuron receives signals from other neurons; the neurons’ inputs are affected by associated weights so that the output signal can be calculated as the output of an activation function. The input is the sum of the neuron’s inputs multiplied by the associated weights. The interconnection of neurons that are grouped into layers allows MNN to approximate nonlinear relationships [61].
Thus, the initial proposal is to use MNN to execute the functions and tasks already mentioned by the ERP and MES systems. Fig. 3.5 shows the approach considering the MNN and the functions mentioned as part of the DT-3 virtual model.
Figure 3.5: DT-3 Virtual Model
Figure 3.5 shows a Multilayer Neural Network (MNN) in which the inputs are Avail- ability and PCA Data; those inputs are represented by only one input neuron, but they are expanded to a cluster of neurons according to the required inputs. On the other hand, the outputs are Production Orders, Finished Goods Waiver, and PCA Data; those outputs also have to be expended to accomplish the required outputs. In general terms, the MNN only
represents the input-output relationship, but it does not represent the total number of neurons required in the input and output layer.
The proposal consists of 3 MNNs that represent the three functions considered for this initial proposal. For practical graphical representation purposes, only the first MNN illustrates a general configuration of the MNN body. In the first MNN, it is observed that the input and output layers contemplate the tasks performed by the order processing (OPR) function. In this case, the OPR function has two input tasks and three output tasks. It is important to mention that the task name reference shown in Fig. 3.5 actually involves all possible values (numeric, characters, etc.) that can be obtained from that specific task. The hidden layers of the network will have a certain number of neurons each associated with an activation function. Remember that an artificial neuronal system tries to model the behavior of a biological network, in MNN the body of each neuron represents a linear adder of external stimuli followed by an activation function (non-linear) whose task is to use the sum of the stimuli to determine the output of the activity of the neuron [62].
Some of the inputs and outputs of each network are labeled ”data”. Taking the first PC network as an example: the inputs and outputs with this label are shown as a single value for illustrative simplicity, but note that these tasks consist of two inputs and two outputs, i.e. the value of ”QA Data” as input is actually made up of two sets of values: ”Standards and Clients” and “QA Results”, depending on the functional control model on which this DT is based. This relationship is applied to the rest of the outputs associated with the PC network labeled as ”Data”. For the OPR network: PCA Data is the information that is shared between the PC and the OPR functions defined in the first part of the standard, in the functional enterprise/control model (see Fig. 2.3 ). In this network, the value of PCA data as input does not necessarily correspond to PCA data information values as output of the network. For more detail on the information that each label represents within this DT-3 model, one can consult the first part of the standard, where the functions carried out in each entity and the information exchanged between them are defined.
On the other hand, some of the tasks in each of the OPR, PS and PC networks have an arrow after or before the label with a certain color. Those with the same color indicate that the set of values derived from this task as the output will be the same set of input values that some of the two other networks will receive. For instance, the PS network output data set for ”Availability” will be the same data set that the OPR network will receive as input.
The graphic union between each network indicates that they are part of the same DT and that the networks will have to be executed as a single system whose internal subsystems interact simultaneously and from which partial and final results are obtained in real time.
The main objective of the application of the MNNs is to have a virtual intelligent system that has the ability to learn and apply that knowledge automatically. Considering this and the nature of the environment where the activities of the input layer are executed and the objective results at the output of the proposed networks, a Back Propagation (BP) algorithm is proposed for the training of the MNN. BP algorithms work under supervised learning, therefore you need a set of training instructions that describe each output and its expected output value. To “train” the neural network, it is necessary to set a data set containing input signals connected with corresponding targets. In each iteration of the training process, the values of the weights are adjusted using new data from the dataset defined for training. Weight modifications are calculated using the backward error algorithm for supervised training. Each
step of the training begins by forcing the inputs out of the training set. Then it is possible to determine the output values of the signals of each neuron in each layer of the network [62].
The DT-3 model has as its initial objective to emulate the decision making that is carried out in typical ERP and MES systems, focusing for now on covering those functions discussed in the development of this work. Through the DT-3 model, anticipated answers can be ob- tained for those tasks that are simulated within any of the proposed networks. This can be a valuable alternative in the simulation of scenarios with conditions attached to reality and responses in a short period of time, or responses scheduled to be displayed at a certain time.
Therefore, through the DT-3 model it will be possible to simulate and have answers for a set of activities such as: production monitoring, providing information to control batches of products and labeling them for identification, obtaining information on material waste , ma- chine downtime and process status data, etc. In summary, the activities related to production control that are specifically contemplated for the PC functions, and all those contemplated in the ISA-95 standard for the OPR and PS functions.
With the results coming out of the DT-3 models, decision-makers can prevent, predict and improve specific points for each of the simulated tasks.
Assessment of a Didactic Manufacturing System through the Virtual AP Model
In the same way that the traditional pyramid can be applied to any manufacturing entity, the DT model proposed in this work can be applied to a company as a whole, to a manufacturing plant, to a designated plant area or to a particular process.
Therefore, this chapter presents the evaluation of a manufacturing system through the model proposed in the previous chapter. The objective of the evaluation is the application of the model as part of the development of DTs for workstations applied in the teaching of undergraduate students. At the end of this case study, opportunity areas of the evaluated sys- tem will be obtained and will be addressed in the next chapter. Thus, in the first instance, the didactic system developed in the Remote Laboratory at the Tec de Monterrey CCM will be described and then its components will be located and discussed within the framework of the proposed model. The didactic manufacturing system was developed by undergraduate students and researchers from Tec de Monterrey CCM, the information presented in this sec- tion was provided as a collaboration for this work, and the analysis carried out here is based exclusively on the information provided and presented in that work.